Speech recognition systems use acoustic models to transcribe utterances. Acoustic models are often trained using a set of training data that include recorded utterances. Typically, acoustic models include statistical data about the sounds that make up utterances in the training data.